- 06 Jan, 2016 40 commits
-
-
Jiri Olsa authored
We can't convert u16 cpu_map_entries::cpu[x] value directly to int, because it could hold -1, which would be converted as 65535. Adding special treatment for -1, which is not real cpu number, to be converted to (int -1). Reported-by:
Arnaldo Carvalho de Melo <acme@kernel.org> Signed-off-by:
Jiri Olsa <jolsa@kernel.org> Cc: David Ahern <dsahern@gmail.com> Cc: Kan Liang <kan.liang@intel.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1452077397-31958-2-git-send-email-jolsa@kernel.orgSigned-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com>
-
Jiri Olsa authored
Add support to get stat events data in perf python scripts. The python script shall implement the following new interface to process stat data: def stat__<event_name>_[<modifier>](cpu, thread, time, val, ena, run): - is called for every stat event for given counter, if user monitors 'cycles,instructions:u" following callbacks should be defined: def stat__cycles(cpu, thread, time, val, ena, run): def stat__instructions_u(cpu, thread, time, val, ena, run): def stat__interval(time): - is called for every interval with its time, in non interval mode it's called after last stat event with total measured time in ns The rest of the current interface stays untouched.. Please check example CPI metrics script in following patch with command line examples in changelogs. Signed-off-by:Kan Liang <kan.liang@intel.com> Cc: David Ahern <dsahern@gmail.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1452028152-26762-8-git-send-email-jolsa@kernel.org [ Rename 'time' parameters to 'tstamp', to fix the build in older distros ] Signed-off-by:
-
Jiri Olsa authored
Implement struct scripting_ops::(process_stat|process_stat_interval) handlers - calling scripting handlers from stat events handlers. Signed-off-by:
-
Jiri Olsa authored
Python and perl scripting code will define those callbacks and get stat data. Signed-off-by:
-
Jiri Olsa authored
Adding processing of stat config event and initialize stat_config object. Signed-off-by:
-
Jiri Olsa authored
Adding processing of cpu/threads maps. Configuring session's evlist with these maps. Signed-off-by:
-
Jiri Olsa authored
For pipe sessions we need to keep sample_type zero, because script's perf_evsel__check_attr is triggered by sample_type != 0, and the check would fail on stat session. I was tempted to keep it zero unconditionally, but the pipe session is sufficient. In perf.data session we are guarded by HEADER_STAT feature. Signed-off-by:
-
Namhyung Kim authored
Signed-off-by:
Namhyung Kim <namhyung@kernel.org> Acked-by:
-
Namhyung Kim authored
When a perf.data file has multiple events, it's likely to be similar (tracepoint) events. In that case, they might have same field name so add all of them to sort keys instead of bailing out. Signed-off-by:
-
Jiri Olsa authored
Using FEATURE-DUMP in bpf subproject for features detection in case bpf is built via perf. Keeping the current features detection otherwise. Signed-off-by:
Wang Nan <wangnan0@huawei.com> Cc: David Ahern <dsahern@gmail.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Zefan Li <lizefan@huawei.com> Cc: pi3orama <pi3orama@163.com> Link: http://lkml.kernel.org/r/1450893514-9158-6-git-send-email-jolsa@kernel.orgSigned-off-by:
-
Jiri Olsa authored
Changing the contents of the FEATURE-DUMP file, so it looks like: feature-backtrace=1 feature-dwarf=0 feature-fortify-source=1 feature-sync-compare-and-swap=0 This way it could get included in sub projects, so they won't be forced to redo features detection. Also now storing the complete set of features. Signed-off-by:
-
Jiri Olsa authored
The feature_assign macro generates feature value assignment for name, like: $(call feature_assign,dwarf) == feature-dwarf=1 This will be used more in following patches. Signed-off-by:
-
Jiri Olsa authored
We decide what dwarf unwind to choose way after the Makefile.feature makefile is included. The $(dwarf-post-unwind) is not even set at that time. For the same reason it was never included in FEATURE-DUMP file. Moving it into perf VF=1 verbose display. $ make VF=1 BUILD: Doing 'make -j4' parallel build Auto-detecting system features: ... dwarf: [ on ] ... ... LIBUNWIND_DIR: ... LIBDW_DIR: ... DWARF post unwind library: libunwind ... Signed-off-by: -
Jiri Olsa authored
This function is cursed.. ;-) Signed-off-by:
-
Namhyung Kim authored
When an evlist contains tracepoint events only, use 'trace' sort key as default. If --raw-trace option was given, use 'trace_fields' instead. This will make users more convenient to see trace result. Suggested-and-Acked-by:
-
Namhyung Kim authored
The 'trace_fields' sort key is similar as 'trace' sort key, but it shows each fields separately. Each event will get different columns as their fields. $ perf report -s trace_fields --stdio # To display the perf.data header info, please use --header/--header-only options. # # # Total Lost Samples: 0 # # Samples: 20K of event 'kmem:kmalloc' # Event count (approx.): 20533 # # Overhead Command call_site ptr bytes_req bytes_alloc gfp_flags # ........ ....... .................. .................. ......... ........... ................... # 99.89% perf ffffffffa01d4396 0xffff8803ffb79720 96 96 GFP_NOFS|GFP_ZERO 0.06% sleep ffffffff8114e1cd 0xffff8803d228a000 4096 4096 GFP_KERNEL 0.03% perf ffffffff811d6ae6 0xffff8803f7678f00 240 256 GFP_KERNEL|GFP_ZERO 0.00% perf ffffffff812263c1 0xffff880406172380 128 128 GFP_KERNEL 0.00% perf ffffffff812264b9 0xffff8803ffac1600 504 512 GFP_KERNEL 0.00% perf ffffffff81226634 0xffff880401dc5280 28 32 GFP_KERNEL 0.00% sleep ffffffff81226da9 0xffff8803ffac3a00 392 512 GFP_KERNEL # Samples: 20K of event 'kmem:kfree' # Event count (approx.): 20597 # # Overhead call_site ptr # ........ .................. .................. # 99.58% ffffffffa01d85ad 0xffff8803ffb79720 0.07% ffffffff81443f5c 0xffff8803f7669400 0.02% ffffffff811d5753 0xffff8803f7678f00 0.01% ffffffff81443f5c 0xffff8803f766be00 0.01% ffffffff8114e359 0xffff8803d228a000 0.01% ffffffff81443f5c 0xffff8800d156dc00 0.01% ffffffff81443f5c 0xffff8803f7669400 0.01% ffffffff8114e359 0xffff8803d228a000 0.01% ffffffff8114e359 0xffff8803d228a000 0.01% ffffffff8114e359 0xffff8803d228a000 Signed-off-by: -
Namhyung Kim authored
When there are multiple events, each dynamic sort key is defined just for one event. In this case other events will always show "N/A" for those fields. But they are meaningless and consume precious screen width. Let's skip those undefined dynamic fields. $ perf record -e kmem:kmalloc,kmem:kfree -a sleep 1 $ perf report -s 'comm,kmalloc.*' --stdio # To display the perf.data header info, please use --header/--header-only options. # # # Total Lost Samples: 0 # # Samples: 20K of event 'kmem:kmalloc' # Event count (approx.): 20533 # # Overhead Command call_site ptr bytes_req bytes_alloc gfp_flags # ........ ....... .................. .................. ......... ........... ................... # 99.89% perf ffffffffa01d4396 0xffff8803ffb79720 96 96 GFP_NOFS|GFP_ZERO 0.06% sleep ffffffff8114e1cd 0xffff8803d228a000 4096 4096 GFP_KERNEL 0.03% perf ffffffff811d6ae6 0xffff8803f7678f00 240 256 GFP_KERNEL|GFP_ZERO 0.00% perf ffffffff812263c1 0xffff880406172380 128 128 GFP_KERNEL 0.00% perf ffffffff812264b9 0xffff8803ffac1600 504 512 GFP_KERNEL 0.00% perf ffffffff81226634 0xffff880401dc5280 28 32 GFP_KERNEL 0.00% sleep ffffffff81226da9 0xffff8803ffac3a00 392 512 GFP_KERNEL # Samples: 20K of event 'kmem:kfree' # Event count (approx.): 20597 # # Overhead Command # ........ .............. # 99.63% perf 0.14% sleep 0.11% irq/36-iwlwifi 0.11% kworker/u16:0 0.01% Xorg 0.00% firefox Signed-off-by: -
Namhyung Kim authored
Support '*' character for field name to add all (non-common) fields as sort keys easily. $ perf report -s 'switch.*' --stdio ... # Overhead prev_comm prev_pid prev_prio prev_state next_comm next_pid next_prio # ........ ........... ......... ......... .......... ............ ........ ......... # 3.82% swapper/0 0 120 0 netctl-auto 18711 120 3.75% netctl-auto 18711 120 1 swapper/0 0 120 2.24% swapper/1 0 120 0 netctl-auto 18709 120 2.24% netctl-auto 18709 120 1 swapper/1 0 120 1.80% swapper/2 0 120 0 rcu_preempt 7 120 1.80% swapper/2 0 120 0 netctl-auto 18711 120 1.80% rcu_preempt 7 120 1 swapper/2 0 120 1.80% netctl-auto 18711 120 1 swapper/2 0 120 ... Suggested-and-acked-by:Jiri Olsa <jolsa@redhat.com> Signed-off-by:
-
Namhyung Kim authored
The dynamic sort key requires event name but specifying full event name is rather inconvenient. This patch adds more ways to identify the event in a more compact way. 1. If session has just one event, event name can be omitted. 2. Events can be accessed by index preceded by a percent sign. 3. A part of the name can be used, if it's not ambiguous. The partial name should not contain ':' in it. 4. Full system + event name is still used, it should contain ':'. So in the below example all does same thing: $ perf record -e sched:sched_switch -a sleep 1 $ perf report -s next_pid,next_comm $ perf report -s %1.next_pid,%1.next_comm $ perf report -s switch.next_pid,switch.next_comm $ perf report -s sched:sched_switch.next_pid,sched:sched_switch.next_comm Signed-off-by: -
Namhyung Kim authored
The --raw-trace option allows disabling pretty printing by the event's print_fmt or plugin. Besides that, each dynamic sort key now can receive a 'raw' suffix separated by '/' to ask for the raw trace of a specific field. $ perf report -s comm,kmem:kmalloc.gfp_flags ... # Overhead Command gfp_flags # ........ ....... ................... # 99.89% perf GFP_NOFS|GFP_ZERO 0.06% sleep GFP_KERNEL 0.03% perf GFP_KERNEL|GFP_ZERO 0.01% perf GFP_KERNEL Now $ perf report -s comm,kmem:kmalloc.gfp_flags --raw-trace or $ perf report -s comm,kmem:kmalloc.gfp_flags/raw ... # Overhead Command gfp_flags # ........ ....... .......... # 99.89% perf 32848 0.06% sleep 208 0.03% perf 32976 0.01% perf 208 Suggested-and-Acked-by: -
Namhyung Kim authored
The 'trace' sort key is to show tracepoint event output using either print fmt or plugin. For example sched_switch event (using plugin) will show output like below: # perf record -e sched:sched_switch -a usleep 10 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.197 MB perf.data (69 samples) ] # $ perf report -s trace --stdio ... # Overhead Trace output # ........ ................................................... # 9.48% swapper/0:0 [120] R ==> transmission-gt:17773 [120] 9.48% transmission-gt:17773 [120] S ==> swapper/0:0 [120] 9.04% swapper/2:0 [120] R ==> transmission-gt:17773 [120] 8.92% transmission-gt:17773 [120] S ==> swapper/2:0 [120] 5.25% swapper/0:0 [120] R ==> kworker/0:1H:109 [100] 5.21% kworker/0:1H:109 [100] S ==> swapper/0:0 [120] 1.78% swapper/3:0 [120] R ==> transmission-gt:17773 [120] 1.78% transmission-gt:17773 [120] S ==> swapper/3:0 [120] 1.53% Xephyr:6524 [120] S ==> swapper/0:0 [120] 1.53% swapper/0:0 [120] R ==> Xephyr:6524 [120] 1.17% swapper/2:0 [120] R ==> irq/33-iwlwifi:233 [49] 1.13% irq/33-iwlwifi:233 [49] S ==> swapper/2:0 [120] Note that the 'trace' sort key works only for tracepoint events. If it's used to other type of events, just "N/A" will be printed. Suggested-and-acked-by: -
Namhyung Kim authored
Each tracepoint event has format string for print to improve readability. Try to parse the output and match the field name. If it finds one, use that for the result. If not, fallbacks to the original output. For example, sort on kmem:kmalloc.gfp_flags looks like below: (Note: libtraceevent plugins are not installed on my system. They might affect the output below) Before: # Overhead Command gfp_flags # ........ ....... .......... # 99.89% perf 32848 0.06% sleep 208 0.03% perf 32976 0.01% perf 208 After: # Overhead Command gfp_flags # ........ ....... ................... # 99.89% perf GFP_NOFS|GFP_ZERO 0.06% sleep GFP_KERNEL 0.03% perf GFP_KERNEL|GFP_ZERO 0.01% perf GFP_KERNEL Signed-off-by: -
Namhyung Kim authored
The existing sort keys are less useful for tracepoint events in that they are always sampled at the same place, the function where the tracepoint is located. For example, a 'perf report' on sched:sched_switch event looks like the following: # Overhead Command Shared Object Symbol # ........ ............... ................ .............. # 47.22% swapper [kernel.vmlinux] [k] __schedule 21.67% transmission-gt [kernel.vmlinux] [k] __schedule 8.23% netctl-auto [kernel.vmlinux] [k] __schedule 5.53% kworker/0:1H [kernel.vmlinux] [k] __schedule 1.98% Xephyr [kernel.vmlinux] [k] __schedule 1.33% irq/33-iwlwifi [kernel.vmlinux] [k] __schedule 1.17% wpa_cli [kernel.vmlinux] [k] __schedule 1.13% rcu_preempt [kernel.vmlinux] [k] __schedule 0.85% ksoftirqd/0 [kernel.vmlinux] [k] __schedule 0.77% Timer [kernel.vmlinux] [k] __schedule In fact, tracepoints have meaningful information in their fields but there's no way to use in 'perf report' currently. The dynamic sort keys are introduced in this patc to overcome this limitation. The sched:sched_switch events have following fields: # sudo cat /sys/kernel/debug/tracing/events/sched/sched_switch/format name: sched_switch ID: 268 format: field:unsigned short common_type; offset:0; size:2; signed:0; field:unsigned char common_flags; offset:2; size:1; signed:0; field:unsigned char common_preempt_count; offset:3; size:1; signed:0; field:int common_pid; offset:4; size:4; signed:1; field:char prev_comm[16]; offset:8; size:16; signed:1; field:pid_t prev_pid; offset:24; size:4; signed:1; field:int prev_prio; offset:28; size:4; signed:1; field:long prev_state; offset:32; size:8; signed:1; field:char next_comm[16]; offset:40; size:16; signed:1; field:pid_t next_pid; offset:56; size:4; signed:1; field:int next_prio; offset:60; size:4; signed:1; print fmt: "prev_comm=%s prev_pid=%d prev_prio=%d prev_state=%s%s ==> next_comm=%s next_pid=%d next_prio=%d", REC->prev_comm, REC->prev_pid, REC->prev_prio, REC->prev_state & (2048-1) ? __print_flags(REC->prev_state & (2048-1), "|", { 1, "S"} , { 2, "D" }, { 4, "T" }, { 8, "t" }, { 16, "Z" }, { 32, "X" }, { 64, "x" }, { 128, "K"}, { 256, "W" }, { 512, "P" }, { 1024, "N" }) : "R", REC->prev_state & 2048 ? "+" : "", REC->next_comm, REC->next_pid, REC->next_prio With dynamic sort keys, you can use <event.field> as a sort key. Those dynamic keys are checked and created on demand. For instance, below is to sort by next_pid field output on the same data file: $ perf report -s comm,sched:sched_switch.next_pid --stdio ... # Overhead Command next_pid # ........ ............... .......... # 21.23% transmission-gt 0 20.86% swapper 17773 6.62% netctl-auto 0 5.25% swapper 109 5.21% kworker/0:1H 0 1.98% Xephyr 0 1.98% swapper 6524 1.98% swapper 27478 1.37% swapper 27476 1.17% swapper 233 Multiple dynamic sort keys are also supported: $ perf report -s comm,sched:sched_switch.next_pid,sched:sched_switch.next_comm --stdio ... # Overhead Command next_pid next_comm # ........ ............... .......... ................ # 20.86% swapper 17773 transmission-gt 9.64% transmission-gt 0 swapper/0 9.16% transmission-gt 0 swapper/2 5.25% swapper 109 kworker/0:1H 5.21% kworker/0:1H 0 swapper/0 2.14% netctl-auto 0 swapper/2 1.98% netctl-auto 0 swapper/0 1.98% swapper 6524 Xephyr 1.98% swapper 27478 netctl-auto 1.78% transmission-gt 0 swapper/3 1.53% Xephyr 0 swapper/0 1.29% netctl-auto 0 swapper/1 1.29% swapper 27476 netctl-auto 1.21% netctl-auto 0 swapper/3 1.17% swapper 233 irq/33-iwlwifi Note that pid 0 exists for each cpu so have comm of 'swapper/N'. Signed-off-by: -
Namhyung Kim authored
This is a preparation to support dynamic sort keys for tracepoint events. Dynamic sort keys can be created for specific fields in trace events so it needs the event information. Signed-off-by:
-
Namhyung Kim authored
This is a preparation to support dynamic sort keys for tracepoint events. Dynamic sort keys can be created for specific fields in trace events so it needs the event information, so we need to pass the evlist to the sort routines, create it sooner so that the next patch can do that. Signed-off-by:
-
Namhyung Kim authored
The print_event_field() and print_event_fields() functions print basic information of a given field or event without the print format. They'll be used by dynamic sort keys later. Committer note: Rename it to pevent_print_field[s]() to get proper namespacing, as discussed with Steven Rostedt. Signed-off-by:
Steven Rostedt <rostedt@goodmis.org> Cc: Andi Kleen <andi@firstfloor.org> Cc: David Ahern <dsahern@gmail.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Wang Nan <wangnan0@huawei.com> Link: http://lkml.kernel.org/r/1450876121-22494-1-git-send-email-namhyung@kernel.orgSigned-off-by:
-
Namhyung Kim authored
The raw_data and raw_size fields are to provide tracepoint specific information. They will be used by dynamic sort keys later. Signed-off-by:
-
Namhyung Kim authored
This is a preparation to add more info into the hist_entry. Also it already passes too many argument, so passing sample directly will reduce the overhead of the function call. Signed-off-by:
-
Vince Weaver authored
This is a long standing bug with the l1-dcache-stores generic event on AMD machines. My perf_event testsuite has been complaining about this for years and I'm finally getting around to trying to get it fixed. The data_cache_refills:system event does not make sense for l1-dcache-stores. Maybe this was a typo and it was meant to be for l1-dcache-store-misses? In any case, the values returned are nowhere near correct for l1-dcache-stores and in fact the umask values for the event have completely changed with fam15h so it makes even less sense than ever. So just remove it. Signed-off-by:
Vince Weaver <vincent.weaver@maine.edu> Signed-off-by:
Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/alpine.DEB.2.20.1512091134350.24311@vincent-weaver-1.umelst.maine.eduSigned-off-by:
Ingo Molnar <mingo@kernel.org>
-
Harish Chegondi authored
Knights Landing uncore performance monitoring (perfmon) is derived from Haswell-EP uncore perfmon with several differences. One notable difference is in PCI device IDs. Knights Landing uses common PCI device ID for multiple instances of an uncore PMU device type. In Haswell-EP, each instance of a PMU device type has a unique device ID. Knights Landing uncore components that have performance monitoring units are UBOX, CHA, EDC, MC, M2PCIe, IRP and PCU. Perfmon registers in EDC, MC, IRP, and M2PCIe reside in the PCIe configuration space. Perfmon registers in UBOX, CHA and PCU are accessed via the MSR interface. For more details, please refer to the public document: https://software.intel.com/sites/default/files/managed/15/8d/IntelXeonPhi%E2%84%A2x200ProcessorPerformanceMonitoringReferenceManual_Volume1_Registers_v0%206.pdfSigned-off-by:
Harish Chegondi <harish.chegondi@intel.com> Signed-off-by:
-
Harish Chegondi authored
Call uncore_pci_box_ctl() function to get the PMON box control MSR offset instead of hard coding the offset. This would allow us to use this snbep_uncore_pci_init_box() function for other PCI PMON devices whose box control MSR offset is different from SNBEP_PCI_PMON_BOX_CTL. Signed-off-by:
-
Harish Chegondi authored
Knights Landing core is based on Silvermont core with several differences. Like Silvermont, Knights Landing has 8 pairs of LBR MSRs. However, the LBR MSRs addresses match those of the Xeon cores' first 8 pairs of LBR MSRs Unlike Silvermont, Knights Landing supports hyperthreading. Knights Landing offcore response events config register mask is different from that of the Silvermont. This patch was developed based on a patch from Andi Kleen. For more details, please refer to the public document: https://software.intel.com/sites/default/files/managed/15/8d/IntelXeonPhi%E2%84%A2x200ProcessorPerformanceMonitoringReferenceManual_Volume1_Registers_v0%206.pdfSigned-off-by:
-
Kan Liang authored
The uncore subsystem for Broadwell-EP is similar to Haswell-EP. There are some differences in pci device IDs, box number and constraints. This patch extends the Broadwell-DE codes to support Broadwell-EP. Signed-off-by:
-
Huang Rui authored
Actually, rapl_sysfs_show is a duplicate of perf_event_sysfs_show. We prefer to use the unified interface. Signed-off-by:
Huang Rui <ray.huang@amd.com> Signed-off-by:
-
Stephane Eranian authored
This patch updates the PEBS support for Intel Atom to provide an alias for the cycles:pp event used by perf record/top by default nowadays. On Atom, only INST_RETIRED:ANY supports PEBS, so we use this event instead with a large cmask to count cycles. Given that Core2 has the same issue, we use the intel_pebs_aliases_core2() function for Atom as well. Signed-off-by:
Stephane Eranian <eranian@google.com> Signed-off-by:
-
Stephane Eranian authored
This patch fixes broken PEBS support on Intel Atom and Core2 due to wrong pointer arithmetic in intel_pmu_drain_pebs_core(). The get_next_pebs_record_by_bit() was called on PEBS format fmt0 which does not use the pebs_record_nhm layout. Signed-off-by:
-
Stephane Eranian authored
This patches fixes the LBR kernel crashes on Intel Atom. The kernel was assuming that if the CPU supports 64-bit format LBR, then it has an LBR_SELECT MSR. Atom uses 64-bit LBR format but does not have LBR_SELECT. That was causing NULL pointer dereferences in a couple of places. Signed-off-by:
-
Stephane Eranian authored
This patch fixes a bug in the filter_events() function. The patch fixes the bug whereby if some mappings did not exist, e.g., STALLED_CYCLES_FRONTEND, then any event after it in the attrs array would disappear from the published list of events in /sys/devices/cpu/events. This could be verified easily on any system post SNB (which do not publish STALLED_CYCLES_FRONTEND): $ ./perf stat -e cycles,ref-cycles true Performance counter stats for 'true': 1,217,348 cycles <not supported> ref-cycles The problem is that in filter_events() there is an assumption that the argument (attrs) is organized in increasing continuous event indexes related to the event_map(). But if we remove the non-supported events by shifing the position in the array, then the lookup x86_pmu.event_map() needs to compensate for it, otherwise we are looking up the wrong index. This patch corrects this problem by compensating for the deleted events and with that ref-cycles reappears (here shown on Haswell): $ perf stat -e ref-cycles,cycles true Performance counter stats for 'true': 4,525,910 ref-cycles 1,064,920 cycles 0.002943888 seconds time elapsed Signed-off-by: -
Andi Kleen authored
Add a new 'three-p' precise level, that uses INST_RETIRED.PREC_DIST as base. The basic mechanism of abusing the inverse cmask to get all cycles works the same as before. PREC_DIST is available on Sandy Bridge or later. It had some problems on Sandy Bridge, so we only use it on IvyBridge and later. I tested it on Broadwell and Skylake. PREC_DIST has special support for avoiding shadow effects, which can give better results compare to UOPS_RETIRED. The drawback is that PREC_DIST can only schedule on counter 1, but that is ok for cycle sampling, as there is normally no need to do multiple cycle sampling runs in parallel. It is still possible to run perf top in parallel, as that doesn't use precise mode. Also of course the multiplexing can still allow parallel operation. :pp stays with the previous event. Example: Sample a loop with 10 sqrt with old cycles:pp 0.14 │10: sqrtps %xmm1,%xmm0 <-------------- 9.13 │ sqrtps %xmm1,%xmm0 11.58 │ sqrtps %xmm1,%xmm0 11.51 │ sqrtps %xmm1,%xmm0 6.27 │ sqrtps %xmm1,%xmm0 10.38 │ sqrtps %xmm1,%xmm0 12.20 │ sqrtps %xmm1,%xmm0 12.74 │ sqrtps %xmm1,%xmm0 5.40 │ sqrtps %xmm1,%xmm0 10.14 │ sqrtps %xmm1,%xmm0 10.51 │ ↑ jmp 10 We expect all 10 sqrt to get roughly the sample number of samples. But you can see that the instruction directly after the JMP is systematically underestimated in the result, due to sampling shadow effects. With the new PREC_DIST based sampling this problem is gone and all instructions show up roughly evenly: 9.51 │10: sqrtps %xmm1,%xmm0 11.74 │ sqrtps %xmm1,%xmm0 11.84 │ sqrtps %xmm1,%xmm0 6.05 │ sqrtps %xmm1,%xmm0 10.46 │ sqrtps %xmm1,%xmm0 12.25 │ sqrtps %xmm1,%xmm0 12.18 │ sqrtps %xmm1,%xmm0 5.26 │ sqrtps %xmm1,%xmm0 10.13 │ sqrtps %xmm1,%xmm0 10.43 │ sqrtps %xmm1,%xmm0 0.16 │ ↑ jmp 10 Even with PREC_DIST there is still sampling skid and the result is not completely even, but systematic shadow effects are significantly reduced. The improvements are mainly expected to make a difference in high IPC code. With low IPC it should be similar. Signed-off-by:
Andi Kleen <ak@linux.intel.com> Signed-off-by:
-
Andi Kleen authored
I added UOPS_RETIRED.ALL by mistake to the Skylake PEBS event list for cycles:pp. But the event is not documented for Skylake, and has some issues. The recommended replacement for cycles:pp is to use INST_RETIRED.ANY+pebs as a base, similar to what CPUs before Sandy Bridge did. This new event is called INST_RETIRED.TOTAL_CYCLES_PS. The event is not really new, but has been already used by perf before Sandy Bridge for the original cycles:p Note the SDM doesn't document that event either, but it's being documented in the latest version of the event list on: https://download.01.org/perfmon/SKL This patch does: - Remove UOPS_RETIRED.ALL from the Skylake PEBS event list - Add INST_RETIRED.ANY to the Skylake PEBS event list, and an table entry to allow cmask=16,inv=1 for cycles:pp - We don't need an extra entry for the base INST_RETIRED event, because it is already covered by the catch-all PEBS table entry. - Switch Skylake to use the Core2 PEBS alias (which is INST_RETIRED.TOTAL_CYCLES_PS) Signed-off-by:
-
